When AI models access your internal tools and data through MCP, security is not optional -- it is foundational. I-Machine builds security into every layer of your MCP infrastructure, from encrypted transport between servers and clients to fine-grained access controls that ensure each AI agent can only reach the resources it is authorized to use. We treat security as a first-class requirement, not an afterthought.
Our role-based access control (RBAC) implementation gives you precise governance over your MCP ecosystem. You define which agents, users, and applications can access which tools and data sources, and our systems enforce those policies consistently across every request.
Audit logging is built into every MCP server we deploy. Every tool invocation, data access, and configuration change is recorded with full context -- who made the request, when it happened, what data was accessed, and what the outcome was. These audit trails are invaluable for compliance reporting, incident investigation, and understanding how AI agents interact with your systems over time.
What we put in place
- One place where access is decided. The agent-side ToolBridge rejects an unauthorized tool call locally, before any network or plugin code runs. The MCP server applies the same policy as the authoritative check, however the client was built.
- Rights granted by tool prefix. Each plugin owns the prefix of its tools, so granting a role access to a plugin is one entry. Roles and their prefixes live in
user_permissions.json, which the server loads at startup and watches for changes. A tool added to an existing plugin needs no role change. - Session rights computed on the server. A session receives the intersection of what the user may call and what the mission template requires. The agent cannot widen either side at runtime.
- An audit trail that is the conversation store itself. Each user message, model response, tool call and tool result is written with a timestamp, keyed by session. A session identifier is enough to replay a run, with the arguments of each tool call and what it returned.
- Installation with no outbound network access. The platform ships as Debian packages from a private APT repository inside your network, with GPG-signed manifests. The signing key stays on the build machine and never reaches a target host. Each package is versioned on its own, and every install or upgrade appears in the standard APT logs.
Key Benefits
- Fine-grained RBAC ensuring each agent accesses only authorized resources
- Comprehensive audit logging for compliance reporting and incident investigation
The same RBAC layer guards our own MCP and REST endpoints, where tool access is filtered per user at the dispatch layer. How that filtering and the conversation store used as an audit trail are implemented is detailed in RBAC and audit logging for MCP.